"There is still no proper GO player in AI that can beat genuine human experts" Eheh.

Now we need this guy talking about AlphaGO

For anyone watching who is unaware, his claim that an AI can't beat a Go Champion is no longer true. Google's AI accomplished this, repeatedly and brutally. It's called AlphaGo.

I love how his Go example has aged. Not holding it against him but i think his thoughts now that we have Go AI would be cool to know.

Neurophysiologist here. Although it is certainly not clear yet how the brain actually does things like play go or keep irrelevant knowledge form interfering with ongoing thought, there are definitely theories about these things that are being tested with increasing success. Saying that it is 'a total mystery' is too much and we have some good candidates. Although I'm personally not an expert on these theories, I think the theory that has the most explanatory power in this case might be idea that the brain works more or less as a dynamical system made up of attractor networks. If you are technically inclined you're probably better of looking it up yourself, but I can try to explain it as best as I understand it. The brain can have an incredible amount of different states; every thought you have or anything that happens subconsciously at any given moment would be one of these states which, altogether, make up a continuous state space. An attractor is a point that the system tends to gravitate towards in response to perturbations. So if you are walking across the street thinking of the weather and you pass a bookstore and you see a book about Patagonian fruit bats, this observation itself moves the state of the brain in the state space, and in doing so it will follow along a continuous attractor, and you will end up in a state which is dominated by thoughts about fruit bats, rather than the weather, which would constitute a different state. In other words; the reason you are not constantly thinking about EVERYTHING that your mind ever stored, is because what is in your mind at a given moment is a state in state space, and you progress from one state to the next along a continuous attractor. If you suddenly find yourself thinking about fruit bats, it's because you either had an external trigger or internally some kind of subconscious association that you were not aware about. You do not just 'randomly' think about fruit bats, even if it might seem so from your conscious perspective. It is important to understand that such a state space can contain states that might be stable, even though they have not been defined individually/discretely. You can imagine things that never happened to you, with limits set by your previous experience. You can imagine to some extent what it would be like to swim in the ocean in the rain, even if you have never done that before. So, I speculate the following, that through association by rigorous training, the state space of a go players mind contains dense attractors for go moves. When the expert sees a go scenario, his/her mind does not consider each move separately like a computer would do, instead his/her line of thought is constrained by previous experience which carved out possibilities in state space that are more likely to lead to victory. There's avenues of possible moves and strategies that the mind of an expert would never wander into simply because they are not along an attractor in state space. In neuroscience there's often quite a bit of discrepancy between the knowledge of experimentalists and theorists, which means that experimentalists like me often don't understand the implications of a theory, so if anyone has a better explanation I think it would be very interesting to hear. But this is one of my working hypotheses that I can see working within the physiological confines of the brain.

So much for his Go-example... :)

Many very interesting points. Thanks.

It's rather extraordinary to think about how quickly AI has moved since this video.

Thunderf00t's dad?

Y'all should redo this interview with him now that an ai has mastered go

I wish he had elaborated more on the subject of the potential for a friendly A.I

Thank you KZfaq for recommending this blast from the past. We haven't reached the singularity yet (I think). For future context, AutoGPT was put out recently and we have a handful of LLMs you can run on your home computer. We still don't have official support for plugins to GPT-4 or the other rumored large AIs from Google and Amazon. It's insane how much difference seven years makes.

The fact that it's only been 1~ year and Google's AI has already beaten the best Go player in the world shows how fast the understanding of AI is advancing now that we have more than enough parallel computing power to run very powerful deep neural networks to solve problems. We can trial-error their structure "bruteforcely" to find the closest to optimum NN layout to solve a specific problem. Improvements in recurrent neural networks like LSTM have also opened the possibility to add memory banks to NNs so that they can process sequential data streams (like natural language / speech / music) and solve problems that rely on time and that rely on context tracking. And this is only the beginning. There is infinite room for improvement and when specific problems like speech recognition / natural language processing get solved, neural network hardware will be built massively to run these programs in any device and further development will step on this hardware and so on. I estimate that within the next 10 years we will have true AI or something most humans won't be able to differentiate from real human beings.

Ironically, a Go AI by Google just super recently beat a Go Master player.

Although we don't know how to make something as complicated as the brain we do know how to grow complex systems from simplistic rule sets. Perhaps we can create an accelerated virtual evolution as apposed to reverse engineering complex biology.

Note that (5:45) three steps makes a factor of 250^3, not (250^2)^2. You multiply by 250 for each additional move, rather than squaring.

Excellent video. I appreciate the balance of long term optimism with short term practicality.

Dr. Holden makes a lot of good points. The only one think he misses the ball on is the one about computational density and the ability to manufacture computer chips in 3 dimensions. The computing power utilized by an AI wouldn't need to fit inside a skull. In fact, it wouldn't even need to fit in a warehouse. The only real limitation on size is how fast you can transfer data from one part to the other, and how much power you can use.

The Law of Accelerating Returns is the best way to understand all this stuff.

I love how he shoves "AI is gonna destroy us all" aside as nonsense, but then is adamant that moore's law is about to end (a notion that, much like 'AI destroying the world', has been going on for decades now, but we still dont see this slowdown)

Excellent video! id love to hear more from Dr Holden!

first thing he said when talking about the 3d brain problem was three D printers with metals and supports made with heat diffusers. And on the topic of "focus" I thought of the brain holding a script that altered the pathways to make specific thoughts quicker

Just discovered this channel. Superb videos. Many thanks.

More videos with this guy please!!! super interesting.

In March of 2016, AlphaGO, a Deep Learning AI beat the world champion in GO in 4 out of 5 games. The first time a computer Go program has beaten a 9-dan professional without handicaps.

I'm wondering the ~10 - 30 people that dislike a computerphile video. As if they clicked it, realized it was actually intellectual, and just said "NO. HUMAN DUMB" and just disliked it out of spite or something

Thanks computerpfile for the compliments :) Im proud of my brain!

It's Thunderfoot's brother.

Increasing seeming like the "smart people can be very wrong when predicting the future" example of Claude Shannon was relevant, but maybe not in the way expected...

Always interesting to see the pace pick up. Go programs have now beaten the world's best human players and I'm writing this just two years after the video was posted by an academic expert in the field.

Regarding the goo-player, once for quite a while ago i was reading a summer course in artificial neural nets, and for project me and my friends choosed to make a ai-player for a game quite similar to go, five an a row, unlimited board size, the idea was a min-max-tree approximated by some kind of self-learning structure, however the problem was kinda too discreet for the ANN we learned about, and the board size kinda gave some problems

Very insightful. Thank you for this video.

Some interesting points, but I'd like to add that the heat issue seems to be dealt with in GPUs at least, as 3D cores are starting to be released soon in the 2016 Pascal chipset. Other manufacturers are certain to follow suit.

Would really like to hear the arguments for why friendly AI is "more likely". In the space of possible minds or possible utility functions the ones compatible with our human values or preferences are a vanishingly small fraction, and I don't see how we can expect ending up with one of those without aiming very, very carefully.

You reinforce my belief that we act, then we create a thesis. Part of human learning is trying to abstract recurrent patterns from that thesis. The more we trust ourselves the less often we consider the thesis consciously and only do it when someone asks for an accounting of your actions or an outcome is not predicted or we do higher level thinking like when considering society and family. Stochastic learning has to work that way.This is why I always listen to myself when I talk because I am just as interested to know what I will say as my audience is.

So you start by saying that experts have been saying that human level AI are close since the 1950s and have been wrong to use that as an argument to say that it's a long way off. Then you casually say that Moore's law is about to end, but that's also something experts have been saying since the 1950s and have been wrong about. Be consistent with your own arguments.

So interesting to watch this now that AlphaGo has beat the best Go players in the world.

More videos with this guy please.

If I had to explain the AI issues in a single sentence, I'd say that we miss a mathematical definition of "intelligence". We can simulate some of the thought processes of intelligent beings, but we still miss the big picture.

Great video (as always) but the title seems almost click-baity as the singularity was quickly brushed-off and friendly AI wasn't discussed at all (or did I miss it)?

loving all the AI stuff recently!

So will the first AI be powered by fusion or will the first AI help to finally get fusion to work?

There's an error in the animation of the amount of work needed to compute the third step out. It shouldn't be (250^2)^2 = 250^4 but instead 250^3. The exponential growth with the branching factor of 250 yields 250^n for the n'th step.

People are tearing this guy up after AlphaGo, but his point is that we are nowhere near making something as complex as the brain which is true. If you took AlphaGo and plugged it into any other scenario it would fail. AlphaGo can't do anything but play Go whereas our brains can adapt and learn so many different complex things

In what sense is chess a done deal though? It is of course clear that barring some very specific types of positions, the AI easily outperforms even the strongest human players, but we have to take into account that humans doesn't really play chess all that well. According to professor Newborn "the science is done" with regard to developing chess-playing AI. Still, chessprograms play very, very far from perfect chess and so the idea that chess has been solved (like checkers is) is a popular, but false one. The average branching-factor of chess is about one tenth that of Go, so it stands to reason that Go will be harder to crack than chess using conventional methods. This comment is just a small addendum trying to clarify a bit on some of the point made in this excellent video.

I learn so much watching these videos

Could you do an episode on Alternating Current? I know it isn't about computers per se, but it is closer than maths or chemistry. Electricity is the life blood of computers, and most people don't even understand the difference between alternating current and direct current, or why we use alternating current for long distance power transmission instead of direct current.

This episode really would benefit from an update!

Moore's Law's failures are similar to the rapid advancements in transportation in the 20th century and the wall that it ran into involving the exotic materials required to create faster planes

4 months later, and we now have AI that can BEAT the best Go players in the world. Proves what he was saying earlier - we have no idea when these breakthroughs happen.

See, I've always thought the singularity was something else, something which seems a lot more imminent. I've always thought the singularity, in a technological sense, was when humans evolved in such a way that we cannot conceive (with our current brains) the things that happen from then on. I've always seen it as us kind of upgrading our species. In the case of computing I've always seen it as some means of interfacing with computers to enhance our natural abilities. Computers are better than us at crunching large sets of very specific data and giving some kind of output. We're good at taking in input and sorting out the relevant bits from the irrelevant bits, but we're not great at keeping those things in context against a larger pool of data. Being connected to an internet kind of information source at the speed of thought, with some assistance in calculation from a processor would be a giant change in the way we think. And I don't think anyone knows what the world would look like if such an interface became smartphone level ubiquitous. So, to me, it's never represented computers trying to be like us, but more just working with us a lot more directly.

2 years on... what are his thoughts on DeepMind’s AlphaGo?

Brilliant, great video, many thanks.

There is a interesting issue here. If we talk about movie style robot AI, i would think we might se them.... kind of, in a near future, but they will be more like T800 then I-robot. Here is the issue. We can make a robot do most things that human do by (relatively simple) agent programs. If we really got a lot of agent program and a good backbone to run them to. We could probably make a robot that act and feals like a human, show emotions but don't have any. I´m a electronic engineering hat due to circumstances have worked with autistic children the last couple of years. One thing is interesting. A lot of people confuses showing emotions with having emotions. Autistic children (and adult for that matter) show there emotions at a very very week level... they do show it, but it might be that week so it´s almost impossible to see. This is quite the opposite sociopat that show fake emotions but don´t really feel them like humans do. A syntetic human analog would work much like a sociopat in that they show fake emotions and stupid people would believe it. But they would not really be that intelligent. They can be program to finish IQ tested with top result, but really never be intelligent. If the program have thousands and thousands of small agent program. Its a little like Neo in the matrix. If the synthetic human don´t know how to play chess, it will not be able to even make the most simple move. But when the chess agent is programmed, it will beat the world champion hands down. And this is true for basically everything. it would not be intelligent, but it would score full score on all intelligent test. It would have no emotions, but show emotions as a world class actor Would not be able to understand the basic concept of chess, still beat the world champion. In the same way they would not really be a competitor for jobs. Jobs that already can be automated will already be that with normal software. And hard manual laubour would already be replaced by machines where no special skills is needed. Most of the hand work sold because it made by hand, a robot hand is not a true hand. There might be a few professions where it actually can be a real competitor. But if we talk about competitor to real work, industrial robots and normal every day software is much more of a real competitor.

I would love to see a video on game AI. I'm not speaking about NPCs that do simple tasks. I'm talking about bots that can adapt to the player.

I think this prof might be one of my favorites

And here we are, just a few years later and a computer is now the world's best go player.

That comment that humans can't always articulate how to do something they are good at is so relatable. I was once asked if I would be interested in writing curriculum for mixing audio, and my response was "No, I don't know how. I just do it"

One fundamental difference - Brains are analogue and not digital. Computers work everything out in 1's and 0's, or "on" and "off" states. Neurons can vary their sensitivity to weak electric fields which basically gives an endless amount of states. Incomprehensible complexity.

7:30 - Studying behavioural science, can confirm we don't understand the brain. Not yet. Lots of progress has been made but there's no complete picture by any stretch. Pretty exciting field right now.

I do love the concept of AI, it can be so beneficial to humanity. It doesn't even have to be human level, just imagine a system that calculates how to distribute the world's resources in the most efficient way possible. No hunger, no wall street.

The argument is not that the doubling will lead to total processing of go. Instead, it suggests we will be able to reach the ability to digitally model a human brain to the level needed to reproduce conscousness. At that point, the machine can play go. Potentially to supermaster levels.

Oopsie on the reliance of Go as an example! How priceless. And in just over 2 years since the upload of this video, Alpha Zero has now cracked the door on the AGI space by demonstrating dominance across 3 different games (including dominance over AlphaGo) all based on a single approach: self-learning neural net. Buckle up kiddies.

This guy's voice is used in Talamasca's song Daydreaming.

i enjoyed this video. i write games and game engines and ignoring data thats not needed for a specific frame (culling) is computationally expensive, yet human brains do that instantly. very very interesting.

Can somebody transcribe the very first sentence of this video? I just can't understand the second part of it.

Even though there are about one trillion omega giga chad braniacs in this comment section waxing about the validity of the speaker. I enjoyed this video even more given our current context. Thank you for exposing these voices, and for doing so YEARS before this became a point of hype and punditry.

I believe the binary system also is limiting in the A.I, as the system tends toward Boolean values, and adding to the 2 dimensional aspect talked about at the end. My hope is for Quantum computers to allow for much more complex decisions and to have their answers observed without programming all of the preconditions, simply by creating the conditions by which the electron is given the question generally and observing it. As you say, it might be right around the corner. I do realize that the electron spin observations are probably something close to binary, in that the polar conditions controlling the spin are programmed and the spin is observed in a very isolated way. However the possibility could still be there that by initiating the spin in a certain way, creates an a.i., and creates one in which more naturally occurring intelligence can be observed.

Does Dr Sean Holden play classical guitar?

What about doing AI processing with quantum processors .How could this improve AI

The trick to a friendly AI is to give it carefully calculated values. Tell it to appreciate human happiness and learning and the phenomenon humans refer to as 'free will'. If it is a true general intelligence that has been designed to value beneficent actions and polite discourse, it will be the best thing that we've ever done. If not, little prevents it from taking a path which may be catastrophic.

The key words are "convergence" and "unification". The farther along we get in our understanding of nature the more unified all the theories of distinct phenomena become. Electricity, magnetism: electromagnetism. Light, gravity: relativity. computing, biology: artificial intelligence. If the human brain can evolve by random mutation and natural selection (given billions of year), then why couldn't human brains then design an equivalent synthetic brain? I imagine that in the near future, self-organizing molecular systems based on amino acids or some homologous polymer will be used to grow 3-dimensional chips that will look something like mushrooms. Instead of material being deposited layer by layer like a 3-d printer it will take in raw material and energy from a stem that will later mature into the output/reader mechanism and grow a cap of computronium/cerebrum. The earliest generations will be almost microscopic but perform on the level of some older PCs, but costing fractions of a penny to grow and requiring only milliwatts of power to run. They'll be able to grow thousands or millions of them on a single pan and they will be able to network together on some kind of grid/stalk. If you can arrange them in a fractal configuration to save space the final super computer might look something like a cauliflower or broccoli, which if you think about it is kind of what all brains look like. The actual design of the intelligence itself would have to be darwinian in nature because coding it and testing it the conventional way would take too long and be too expensive. Some reverse engineering of animal brains would need to take place to discover how single utility functions like 'reproduction' give rise to all the complex behaviors and cognitive abilities and alter it to something like 'obedience' or 'human welfare' or 'free agency of sapient beings' or 'voluntaryism/non-aggression principle' etc. then it's a matter of nudging the emergent behaviors into specializations. Perhaps the way their robot body is designed might inform how their brains grow, adapt, and specialize if that is how animal brains grow adapt and specialize. Biochemists have already created a fully synthetic living cell. They can make molecular machines though with limited functionality. They can model proteins and enzymes in quantum computers for reactivity with drugs. I think the problem with talking about artificial intelligence one to one is that the rise of AI will not be brought about by one individual scientist/engineer/technician. It will be a collaboration between hundreds or thousands.

I love how this was a month before alphago finally did it, 6 months before lee sedol.

Couldn't the computer store related data in the same memory space to shorten the amount of time spent on exhaustively looking for it?

That is so true the human brain can filter out everything without having to go through it 1 by 1 to get to the bit of information required to explain a butterflies migration path and computers have to go through all the data as fast as possible to get to the bit it needs.

Is that a copy of Riley Hobson and Bence on the shelf? :D

Great video. I will pass this along to those who've watched "The Matrix" 100 times too many.

I think having a hardware program (jup, a base program completly on circuits which can't be changed by the computer itself, very oldschool) emulating instincts, etc. would really help. Humans are born with a certain "preset" and are not a completly unwritten hard disk. It doesn't seem like that scientists researching on this topic are thinking in that way.

Alphago won against top Go player in first matchMARCH 2016......remember this date ......remember this date....remember this date :)

Maybe the this situation will revolve other way around. People figure out how to reliably grow neural tissue so that it performs wanted tasks better than chips based on modern design philosophies.

We need an update to this

Thunderfoot's long lost brother?

What if we used learning algorithms and record thousands of matches in go and let powerfull computer try to figure out patterns and then on basis of this patters adapt to new patterns during games it plays?

could an AI not replicate the ability to not actively hold all information by use of tags? It determines what subjects (and thus tags) are being talked about (somehow), dumping what isn't needed from quick access memories, and then loading what is tagged correctly? And if there's excess space, have room for predicted branch tags (off-topic tags, but related or correlated to be within a conversation with the current tags). Perhaps the best solution is to just leave the excess space open for quick writing instead of trying to think ahead.

You can't decouple intelligence from emotion because they grow together. Saying expertise, memories, and "everything" gets filtered out is the wrong way to think about it because it requires energy to maintain those connections. They are active at the neuronal level. You can see the manifestations of someones history on in their posture, these are stereotypes. New membrane (connections) are made your whole life but the volume of the brain shrinks after early 20's, losing connections (white matter) and strengthening others - this is where I think wisdom comes from.

this aged like fine milk...

Moore's Law is either slowing down or we are at the end of an s-curve. I'm an optimist so I'll say the latter. As Feynman said, there's plenty of room at the bottom.

I think that the most obvious (and perhaps efficient) way to reach human level AI is to start from the Natural Intelligence we know , ours. Neuroscience is advancing really fast (knowledge doubles ~ every 5 years) , understand and emulate ; that's how the humanity made his / her biggest archivments. Sure emulation is not the only path , but neither is Artificial intelligence , Intelligence Augmentation is one too.

how far away until we start custom growing brains?

Could someone please put captions on this ?

im not a computer scientist, or a neurologist.... but maybe its not the size or ammount of micro processers. maybe its the "wiring" like.. neurons in a brain dont just follow circuits that run through a logic gate. they kind of connect to each other. and they can form new connections. so iit might sound crazy but maybe we could try something like that? or perhaps its the material? rather than copper and silicon, could we make a "computer" out of brain? an organic computer? plus our circuits are usually flat. couldnt we make a bigger proccessor in terms of space then just in terms of making the components smaller. like... make a 3d procesor with more room for circuits? as i said im not an expert, just a thought i had.

So... the game of Go has been beaten by a computer this very week. Should we be worried now?

Very intelligent and solid points all around. I believe that with current technology we could create AI, but we do not have anywhere near enough understanding of consciousness. It would start off a very clumsy thing, probably the size of several plane silos, still very sophisticated, just clumsy compared to the complexity and interconnectedness of the human brain. Either way though, we absolutely should not be working in that direction. Additionally, I have always wondered if the reason why we can filter so effectively and process information so quickly is because we use super fast but lossy processes. Most of your childhood memories are fake, we ALL have those moments when we walk into a room and forget why we are there, what about the word on the tip of your tongue?

Watching this in 2024 right after the announcement of gpt 4o. I was looking for something to tell me that singularity is far away and i found this. I'm terrified by how wrong he is. I'm sure he's a smart guy but this only shows how little we understand about the future and how rapidly the technology is changing. Really, the singularity is near

-Buff -Leather jacket -Interesting facial piercings This guy is definitely a Matrix program.

I'm of the opinion that artificial intelligence will come from the prosthesis sector. If you make nanites that can also do the function of neurons, then give them to someone with some disorder of the nervous system, what will happen if the nanite neurons are able to keep functioning, even after the original neurons have died? If someone has a bundle of nanites that serves all the purposes of their original brain instead of a meat brain, would they qualify as an artificial intelligence? If one tells the nanites to provide a snapshot, and said snapshot is compared to other snapshots, would we not be able to manufacture snapshots of our own through reverse engineering and fiddling with bits of the snapshot? It's like the boat repair thought experiment or the axe one. All the component parts getting replaced over time. Is it the same as the original after the repairs?

Human level Ai is a real deal, i hope you guys finally solve control problem

So I heard graphite processors would become the next "big thing", can anyone explain why that didn't work out anyway?

that is why ai is closer than we think. the estimates are based on the number of neurons and connections and the processing equivalence. consciousness and intelligence does not require all of it... just the right algorithm

I actually LOLed when he made that Go example